When an agent works through a task, if any piece of that task requires editing an already created artifact (implementation plan, task, etc). It stops and says it requires input. However, there is nothing to approve and so the agent cannot do anything else. Essentially each agent can only answer a single prompt, and then has to be shut down (because the agents try to update task/implementation/walkthrough artifacts as part of future responses).
Using: Gemini 3.1 (low) in planning mode
Windows 11
This is only a problem for 3.1 models. Gemini Flash and Claude models have no issues interacting with artifacts.
This means that the 3.1 models are completely useless with Antigravity.
Nevermind, affects all models.
Submitted bug report via antigravity interface, however it apparently doesn’t give a confirmation that report was submitted anymore, so no idea if it actually was sent in.
Who the heck does google have designing the UI/UX for antigravity? Cause they are absolutely atrocious.
Oh wait, it’s probably gemini, that would explain the horrid design choices.
1 Like
Based on my decade of experience with AppSheet (a Google acquisition), I can say pretty confidently that Google does not embrace QA at all. All the testing they do is the developers “testing” their own work, and end-users reporting bugs.